7.9 419 Generalized Reduced Gradient Method we find, at X 1, • ffi _ _ _f ffi __x 1 ffi ffi ffi _f ffi Yf = • ffi ffi _ •x2 ff •f fi X1 Zf •x3 = ff = X1 = {Š4(x 0} [D] = 1 32 [D] ,] GR = _ = Yf 2 Š1 [C] = Š1 Š [[D] ff Š9 2. fi 9 2. Šx 3) 3 X1 Š ff 1 32 fi 3 = ff Š9 2. fi 9 2. = • •g 1 •g1 • = [5 •x1 •x2 X1 • •g 1 • = [32] •x3 X1 [C] = DŠ1 = [ 2 (Š2.6 Š 2) Š 2(Š2.6 Š 2) + 4(2 Š 2) Š10.4] [5 Š10.4] = [0.15625 Š0.325] [C]] T_ Zf fi 0 ( ) Š .15625 0.325 =0 ff Š9 2. fi 9 2. are not zero, the point X1 Step 3: Since the components of GR is not optimum, and hence we go to step 4. Step 4: We use the steepest descent method and take the search direction as S = ŠGR = ff 9 fi Š92.2. Step 5: We find the optimal step length along S. (a) Considering the design variables, we use Eq. (7.111) to obtain For y _= For y 2: 2 3 Š (Š2.6) 9 2. =x 1: = 0 .6087 =x Š 3 Š (2) _= Thus the smaller value gives _ T = Š([D Š1] Š9 2. 1 For z 1 = x 2 : _3 = = 1.1293. = 0 .5435 = 0.5435. Equation (7.113) gives [C])S = Š(0.15625 Š0.325) and hence Eq. (7.114) leads to Thus _ 1 Š 3 Š (2) Š 4.4275 = 1 .1293 ff 9 fi = Š4.4275 Š92.2. 420 Nonlinear Programming III: Constrained Optimization Techniques (b) The upper bound on _ is given by the smaller of _ to 0.5435. By expressing 1 and _ 2, which is equal ff Y + _S fi X= Z + _T we obtain _ _ _ _ 9 _ _Š2.6 + 9.2_ _ _Š2.6 _ _x1 _ _ 2. 2 Š 9.2_ 2 +_ = x2 = X= 2. Š 4.4275 _ x3 _ 2 _ 2 Š 4.4275_ _ Š9 and hence f (_) = f (X) = (Š2.6 + 9.2_ Š 2 + 9.2_)2 + ( 2 Š 9.2_ Š 2 + 4.4275_) 4 = 518.7806_ 4 2 169.28_ + 21.16 + 338.56_ Š df/d_ = 0 gives 2075.1225_ from which we find the root as _ bound value 0.5435, we use _ _. (c) The new vector Xnew Xnew = = + 677.12_ Š 169.28 = 0 _ _ 0.22. Since _ _ is less than the upper is given by . Yold + dY / Zold + dZ _ _ _ ffi Š0.576 ffi ffi Š2.6 + 0.22(9.2) ffi _ ffi ffi ffi ffi = _ 2+ 0 = _ Š 0.024 _ +_ T .22(Š9.2) ffi 2 + 0 .22(Š4.4275) ffi ffi 1 ffi ffi ffi ffi .02595 ffi _ _ . Y +_ old Zold with 3 _S/ dY = ff fi 2 , .024 Š 2.024 dZ = {Š0.97405} Now, we need to check whether this vector is feasible. Since g1 (X new) = (Š0.576)[1 + (Š0.024 2) ] + (1.02595) Š4 3 = Š2.4684 _= 0 the vector Xnew is infeasible. Hence we hold Ynew constant and modify Znew using Newton's method [Eq. (7.108)] as dZ = [D Š] 1 [Šg(X) Š [C]dY] 7.9 Generalized Reduced Gradient Method 421 Since • •g 1 • •z1 [D] = = [4x 3] 3 = [4(1.02595) ]3 = [4.319551] g1 (X) = {Š2.4684} [C] = • •g •y _g 1 • 1 1 = {[2(Š0.576 + 0.024)][Š2(Š0.576 + 0.024) 2 _y + 4 (Š0.024 Š 1.02595) 3]} = [Š1.104 Š3.5258] dZ = × we have Znew becomes •2 1 4 .319551 .4684 Š {Š1.104 Š3.5258} ff 2 fi _ = {Š0.5633} Š.024 2.024 = Zold + dZ = {2 Š 0.5633} = {1.4367}. The current Xnew Xnew = _ _Š0.576 _ Š 0.024 = _ 1 .4367 . Y + dY / old Zold + dZ The constraint becomes g1 = (Š0.576)(1Š(Š0.024 2) 4 3 = 0.6842 _= 0 ) + (1.4367) Š Since this is infeasible, we need to apply Newton's method Xnew (7.108)] at the current X [Eq. . In the present case, instead of repeating new Newton's iteration, we can find the value of Znew = {x 3} new by satisfying the constraint as g 1( X) = (Š0.576)[1 + (Š0.024 or 2) 0 .25 x3 = ( 2.4237) ]+x4 Š3=0 3 = 1 .2477 This gives _ _Š0.576 _ Š 0.024 and Xnew = _ 1 .2477 ) = (Š0.576 + 0.024 2) + (Š0.024 Š 1.2477) f (X new =4 2.9201 Next we go to step 1. Step 1: We do not have to change the set of independent and dependent variables and hence we go to the next step. 422 Nonlinear Programming III: Constrained Optimization Techniques Step 2: We compute the GRG at the current X using Eq. (7.105). Since • ffi _ _f ffi •x1 Yf = _ ffi _f ffi _ •x2 _ ffi ff 2 (Š0.576 + 0.024) ffi = Š 2(Š0.576 + 0.024) + 4(Š0.024 Š 1.2477) ffi ffi Š 1.104 = ff Š 7.1225 • Zf = [C] = fi ff •f fi ff •f fi = {Š4(Š0.024 Š 1.2477 = •z •x3 • •g 11 •g1 • •x1 •x2 = [1.000576 [D] = [D Š] 1[C] GR • •g 1 • •x3 = [4x = [(1 + (Š0.024 ) 2) 3) 0 .027648] 3] 3 = [4(1.2477) 3] = [7.7694] 1 [1.000576 0 .027648] = [0.128784 7 .7694 = _ Yf Š [[D] Š1 [C]] T _ Zf ff Š 1.104 fi Š 7.1225 } = {8.2265} 2 (Š0.576)(Š0.024)] = = fi 3 Š 0 .003558] ff 0 .128784fi ff Š 2.1634 fi ( 8.2265) = Š 7.1518 0 .003558 Since GR _= 0, we need to proceed to the next step. Note: It can be seen that the value of the objective function reduced from an initial value of 21.16 to 2.9201 in one iteration. 7.10 SEQUENTIAL QUADRATIC PROGRAMMING The sequential quadratic programming is one of the most recently developed and perhaps one of the best methods of optimization. The method has a theoretical basis that is related to (1) the solution of a set of nonlinear equations using Newton's method, and (2) the derivation of simultaneous nonlinear equations using Kuhn-Tucker conditions to the Lagrangian of the constrained optimization problem. In this section we present both the derivation of the equations and the solution procedure of the sequential quadratic programming approach. 7.10.1 Derivation Consider a nonlinear optimization problem with only equality constraints: Find X which minimizes f (X) subject to hk (X) = 0, k = 1, 2, . . . , p (7.117)